MSKD: Structured knowledge distillation for efficient medical image segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models have advanced medical image segmentation but are often computationally intensive for clinical use.
• Efficient and accurate medical image segmentation is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To develop an efficient structured knowledge distillation framework for training lightweight medical image segmentation networks.
• To improve the practical implementation of deep learning in clinical settings.

Main Methods:

• Proposed Feature Filtering Distillation to transfer region-level semantic information and minimize redundant data.
• Introduced Region Graph Distillation to leverage graph capabilities for imitating structured semantic information.
• Validated methods on Synapse multi-organ and KiTS kidney tumor segmentation datasets.

Main Results:

• Significant improvements in segmentation performance for lightweight neural networks, up to 18.56% in Dice coefficient.
• Achieved performance gains without increasing model parameters.
• Demonstrated effectiveness in mitigating segmentation inaccuracies due to similar organ characteristics.

Conclusions:

• The proposed knowledge distillation methods offer an efficient solution for medical image segmentation.
• These techniques enhance the accuracy and practicality of deep learning models in clinical environments.
• Improved segmentation accuracy supports better medical diagnoses and patient treatment.